Development and External Validation of a Machine Learning-Based System for Predicting 4-Year Incident Sarcopenia in Multimorbid Older Adults: Results From Two Prospective Cohorts.

AIM: Sarcopenia is closely linked to multimorbidity in older adults, yet its risk factors remain inadequately defined. Current screening tools also lack sensitivity and dynamic risk assessment. This study aimed to develop and validate a machine learning (ML)-based prediction system for estimating 4-year incident sarcopenia risk in older adults with multimorbidity.

METHODS: We analyzed data from 1430 participants (aged ≥ 60 years, with multimorbidity and no baseline sarcopenia) from the China Health and Retirement Longitudinal Study (CHARLS), splitting them into training (70%) and testing (30%) sets. External validation used 1715 participants from the Health and Retirement Study (HRS). Among the 14 candidate predictors initially identified from the literature, 10 key predictors were selected via LASSO regression. Eight ML models were evaluated using Receiver Operating Characteristic-Area Under the Curve (ROC-AUC), precision-recall curves, calibration, and decision curve analysis, with SHapley Additive exPlanations (SHAP) values enhancing interpretability. A web-based prediction system was developed.

RESULTS: The Random Forest model performed best, achieving an ROC-AUC of 0.952 and accuracy of 0.861 in the training set, plus high specificity (0.635) in validation. SHAP analysis identified BMI (< 22.47 or > 34.33 kg/m) and age (> 66.89 years) as critical risk thresholds. Activities of daily living impairment, depressive symptoms, and female gender increased risk, while drinking behavior and married status were protective. The system enables accurate, interpretable, and dynamic sarcopenia risk assessment.

CONCLUSIONS: The ML-based prediction system addresses the limitations of current screening methods and shows potential for personalized clinical decision-making. Broader validation could further strengthen its clinical applicability.